LightOn Open-Sources 149M Retrieval Models, Setting New BEIR SOTA
_reachsumit · x · 2026-07-30
LightOn has released a fully open-source recipe for training retrieval models, aiming to bridge the reproducibility gap caused by the reliance on closed training data in current SOTA models.
- Data & Models: Researchers reconstructed 665M English contrastive pre-training pairs from 34 public sources and built a fine-tuning dataset with mined hard negatives. This yielded two 149M-parameter models: DenseOn (single-vector dense) and LateOn (ColBERT-style late-interaction).
- Performance: These lightweight models achieve 56.20 and 57.22 average nDCG@10 on the BEIR benchmark, setting new state-of-the-art results for their size class.
- Multilingual Generalization: By translating the English data into 8 languages, the team trained 307M-parameter multilingual variants. Despite sharing a backbone, the dense model degrades on languages outside its translate-train support, whereas the late-interaction model generalizes better to unseen languages through token-level matching.
Related event: LightOn Open-Sources Multilingual Retrieval Models mDenseOn and mLateOn(18 posts)→
More from Research
- SimpleMark: multi-bit LLM watermark bounds per-token distortion with f-divergence budget — rishabh16_ · 2026-10-16
- Stanford paper: generative representation learning cuts tokenizer training compute by 7-8x — HirokatuKataoka · 2026-10-16
- Cohere Labs to present confidence estimation for reasoning models at EMNLP 2026 — Cohere_Labs · 2026-10-16
- Vals RSI Index extrapolates Google models hitting 130% autonomous AI R&D by Oct 2027 — HaydnBelfield · 2026-10-16
- SpatialClaw Pulled Into Agentic Vision Data; Filtering and Evals Were the Real Lift — CMHungSteven · 2026-10-16
- Visual Lean: drag-and-drop GUI for the Natural Numbers Game runs Lean 4 in WASM — MilesCranmer · 2026-10-16